Changelog • maq

All notable changes to maq will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[0.4.0] - 2024-04-18

Add get_aipw_scores for calculating AIPW scores given user-supplied estimates, and add some minor polish to get_ipw_scores. #72, #71

Draw a horizontal line when constructing a new plot depending on user-supplied xlim. #73
Fix a minor discrepancy in integrated_difference. #48

Fix a bug in integrated_difference where the AUC measure is wrong if \bar B exceeds the point at which the curve plateaus. #44

Change the maq function signature to make budget an optional argument. The default behavior (budget = NULL) is to fit the Qini curve up to a maximum spend/unit where each unit that is expected to benefit, is treated. #41

Add integrated_difference(object.lhs, object.rhs, spend) for estimating the area between two Qini curves up to some maximum budget spend. #42

Add a type = c("matrix", "vector") option to predict.maq, optionally returning predictions in the set {0, 1, …, K} if type = "vector". #29
Add a convenience function get_ipw_scores to construct evaluation scores via IPW. #28
Add more documentation on statistical details from the paper. #27

Have predict.maq return a standard dense matrix, and remove dependence on the sparse Matrix package. #30

Fix horizontal.line in plot, the option for whether a curve added to the plot should extend all the way to the right of a main plot, if the added Qini curve’s spend path stops before the spend path of the main plot.

First CRAN beta release (this changelog tracks the R package). The R package currently supports

Fitting Qini curves for an arbitrary number of arms using maq(cate.hat, cost.hat, max.budget, Y.eval, ...), where cate.hat and cost.hat are CATE and cost estimates obtained via some function learned on a training set, Y.eval are evaluation scores on a test set (for example inverse-propensity weighted outcomes) and max.budget is the maximum mean budget/unit to fit the curve on. Setting the option target.with.covariates to FALSE yields a baseline Qini curve that can be used to assess the value of treatment targeting based on covariates.
Computing point estimates and standard errors for any point on the curve with average_gain() (supporting clustered standard errors if fit with clusters).
Comparing arbitrary points on different curves using difference_gain(), yielding standard errors that account for the correlation arising from curves fit on the same evaluation data.
Retrieving the underlying “induced” policy at arbitrary spend levels with predict().
Basic plotting functionality with plot().
Retrieving the full gain/spend/allocation path with summary(maq.object).

The Python bindings (source install only) currently supports

Fitting the Qini curve, and retrieving standard errors for any point on the curve, as well as the implied policy \pi (remaining functionality can be added on the wrapper side - the heavy lifting is done in core/C++, and the input/output is language agnostic).